Vol 443 | 19 October 2006 | doi:10.1038/nature05233
LETTERS
A linguistic model for the rational design of
antimicrobial peptides
Christopher Loose1*, Kyle Jensen1,2,3*, Isidore Rigoutsos1,4 & Gregory Stephanopoulos1
Antimicrobial peptides (AmPs) are small proteins that are used by
the innate immune system to combat bacterial infection in multicellular eukaryotes1. There is mounting evidence that these peptides are less susceptible to bacterial resistance than traditional
antibiotics and could form the basis for a new class of therapeutic
agents2. Here we report the rational design of new AmPs that show
limited homology to naturally occurring proteins but have strong
bacteriostatic activity against several species of bacteria, including
Staphylococcus aureus and Bacillus anthracis. These peptides were
designed using a linguistic model of natural AmPs: we treated the
amino-acid sequences of natural AmPs as a formal language and
built a set of regular grammars to describe this language. We used
this set of grammars to create new, unnatural AmP sequences. Our
peptides conform to the formal syntax of natural antimicrobial
peptides but populate a previously unexplored region of protein
sequence space.
AmPs are ubiquitous among multicellular eukaryotes and are
found in diverse contexts including frog skin, scorpion venom and
human sweat. Recent studies show that some AmPs are active against
pathogens that are resistant to traditional antibiotics such as penicillin, tetracycline and vancomycin3. In addition to their antibiotic uses,
AmPs might have other interesting clinical applications: they act as
adjuvants for the adaptive immune system4 and might be useful in
treating certain cancers5.
The many disease-relevant actions of AmPs are a consequence of
their broad ability to distinguish eukaryotic cells from pathogenic
invaders. In general, AmPs have a net positive charge and an amphipathic three-dimensional structure that gives the peptides an electrostatic affinity to the outer leaflet of the microbial membrane6. This
affinity leads to binding, disruption of the membrane and, ultimately,
microbial cell death7.
Our preliminary studies of natural AmPs indicated that their
amphipathic structure gives rise to a modularity among the different
AmP amino-acid sequences. The repeated usage of sequence modules—which might be a relic of evolutionary divergence and radiation—is reminiscent of phrases in a natural language, such as
English. For example, the pattern QxEAGxLxKxxK (where ‘x’ is
any amino acid) is found in more than 90% of the insect AmPs
known as cecropins. On the basis of this observation, we modelled
the AmP sequences as a formal language—a set of sentences using
words from a fixed vocabulary. In this case, the vocabulary is the set
of naturally occurring amino acids, represented by their one-letter
symbols8.
We conjectured that the ‘language of AmPs’ could be described
by a set of regular grammars. Regular grammars are, in essence,
simple rules that describe the allowed arrangements of words.
These grammars, such as the cecropin pattern mentioned previously,
are commonly written as regular expressions and are widely used to
describe patterns in nucleotide and amino-acid sequences9,10.
To find a set of regular grammars to describe AmPs we used the
Teiresias pattern discovery tool11. With Teiresias, we derived a set of
684 regular grammars that occur commonly in 526 well-characterized eukaryotic AmP sequences from the Antimicrobial Peptide
Database (APD)12 (see Methods). Together, these ,700 grammars
describe the ‘language’ of the AmP sequences. In this linguistic metaphor, the peptide sequences are analogous to sentences and the individual amino acids are analogous to the words in a sentence. Each
grammar describes a common arrangement of amino acids, similar
to popular phrases in English. For example, the frog AmP brevinin1E contains the amino-acid sequence fragment PKIFCKITRK, which
matches the grammar P[KAYS][ILN][FGI]C[KPSA][IV][TS][RKC]
[KR] from our database (the bracketed expression [KAYS] indicates
that, at the second position in the grammar, lysine, alanine, tyrosine
or serine is equally acceptable). On the basis of this match, we would
say that the brevinin-1E fragment is ‘grammatical’.
By design, each grammar in this set of ,700 grammars is ten
amino-acids long and is specific to AmPs—at least 80% of the matches
for each grammar in Swiss-Prot/TrEMBL13 (the APD is a subset of
Swiss-Prot/TrEMBL) are found in peptides annotated as AmPs.
To design unnatural AmPs, we combinatorially enumerated all
grammatical sequences of length twenty (see Methods) in which each
window of size ten in the 20-mers was matched by one of the ,700
grammars. We chose this length because we could easily chemically
synthesize 20-mers and this length is close to the median length of
AmPs in the APD. From this set, we removed any 20-mers that had
six or more amino acids in a row in common with a naturally occurring AmP; we then clustered the remaining sequences on the basis
of similarity, choosing 42 to synthesize and assay for antimicrobial
activity. This process is illustrated in Fig. 1 and the designed
sequences are shown as Supplementary Table 1.
For each of the 42 synthetic peptides, we also designed a shuffled
sequence, in which the order of the amino acids was rearranged
randomly so that the sequence did not match any grammars. These
shuffled peptides had the same amino-acid composition as their
synthetic counterparts and thus, the same molecular weight, charge
and isoelectric point (bulk physiochemical factors that are often
correlated with antimicrobial activity). We hypothesized that
because the shuffled sequences were ‘ungrammatical’ they would
have no antimicrobial activity, despite having the same bulk physiochemical characteristics. In addition, we selected eight peptides from
the APD as positive controls and six 20-mers from non-antimicrobial
proteins as negative controls. A complete list of designed and shuffled
sequences, along with controls, is provided in the Supplementary
Information.
1
Department of Chemical Engineering, Massachusetts Institute of Technology, Cambridge, Massachusetts 02139, USA. 2Harvard–-MIT Health Sciences and Technology, Cambridge,
Massachusetts 02139, USA. 3Agrivida, 411 Massachusetts Ave B1, Cambridge, Massachusetts 02139, USA. 4IBM Research Division, Thomas J.Watson Research Center, Yorktown
Heights, New York 10598, USA.
*These authors contributed equally to this work.
867
©2006 Nature Publishing Group
LETTERS
NATURE | Vol 443 | 19 October 2006
Natural
AmPs
~700
grammars
K[QR]QTHA
R[LM]GAG
RLLM[RQ]QTHA
K[QR]QTHA
K[QR]QTHA
K[QR]Q
QRT
Designed sequences of
overlapping grammars
T
T
T
I A
L
L
LL
I LL
G
G
ST
P
P
Scored population of
designed sequences
Experimentally tested
designed sequences
Grammar 1
MKSTGPLLSALLLAVTAGGDP
Grammar 2
S
P
A
G
I
I F V G
FL LA
A
L
VAA S
L
D
T
Q
Filtered to reduce
redundacy and similarity
to natural AmPs
Figure 1 | A schematic of the in silico peptide design strategy. Grammars
are induced from the set of natural AmP sequences using Teiresias.
Overlapping grammars are stitched together to create new 20-amino-acid
sequences that correspond to the antimicrobial syntax. Designed sequences
are scored on the basis of the prevalence of the underlying grammar in
natural peptides. After removing designs with six or more consecutive amino
acids in common with a natural peptide, a representative set of sequences is
selected for experimental validation.
We characterized the activity of each synthetic AmP using a broth
microdilution assay described elsewhere14. This assay measures the
minimum inhibitory concentration (MIC) at which the peptide
inhibits growth of the target organism. Table 1 shows the MICs of
synthetic peptides against Bacillus cereus and Escherichia coli, as representative gram-positive and gram-negative bacteria. Of the 40 soluble (two of the designed and four of the shuffled peptides were
insoluble) designed peptides, 18 had an MIC of 256 mg ml–1 or less
against at least one of the bacterial targets. Only two of the soluble,
shuffled peptides showed antibacterial activity. Thus, the activity is
not an artefact of molecular weight, charge or isoelectric point.
Of the negative controls—six peptides randomly selected from the
middle of non-antimicrobial proteins from Swiss-Prot/TrEMBL—
none had antimicrobial activity, whereas six of the eight naturallyoccurring AmPs in the positive control group did.
To validate MIC determinations, we measured optical density at
varying concentrations of a representative set of designed, shuffled,
and natural peptides. Supplementary Fig. 1 shows that peptides
inhibit growth at the MICs determined by visual inspection. In addition, plating of these samples confirmed that these peptide-treated
bacterial populations were generally not viable at concentrations of
peptides above the MIC.
Two of the designed peptides, D28 (FLGVVFKLASKVFPAVFGKV)
and D51 (FLFRVASKVFPALIGKFKKK), inhibited B. cereus growth
at 16 mg ml–1, which is close to the MICs of the strong positive
controls melittin and cecropin-melittin hybrid (8 mg ml–1). Peptides
with activity against gram-positive bacteria are particularly exciting
because of the prevalence of drug-resistant nosocomial S. aureus
and the threat of bioterror agents such as B. anthracis, or anthrax.
Therefore, we assayed seven designed peptides that had such activity,
including the highly active D28 and D51 peptides, against the Smith
diffuse strain of S. aureus and the Sterne strain of B. anthracis. As
shown in Table 2, all seven peptides were active against both bacteria
at 256 mg ml–1 or less, whereas only one of the seven shuffled controls
was active (the only active shuffled peptide had MICs of 128 mg ml–1
against S. aureus and 256 mg ml–1 against B. anthracis). Moreover, two
designed peptides, D28 and D51, had MICs of 16 mg ml–1 against
Bacillus anthracis, which is equivalent to the activity of Cecropinmelittin hybrid, a strong natural AmP. D28 also had an MIC of
8 mg ml–1 against S. aureus.
In an attempt to improve the antimicrobial activity of our designed
peptides, we optimized our best candidate, peptide D28, using a
heuristic approach. We created 44 variants of D28 by introducing
mutations that were selected to increase positive charge, increase
hydrophobicity, remove an interior proline residue, or improve
segregation of positive and hydrophobic residues based on a helical
projection. As shown in Supplementary Table 2, 18 of the 44 D28
variants showed improved activity against E. coli, B. cereus or S.
aureus. Many of the D28 variants with improved activity against B.
cereus included a mutation at an internal proline, either to lysine or
glycine. D28 and six of its variants were assayed for bactericidal
activity, and all had activity within a twofold dilution of their MIC.
One variant (R8) had MICs of 16 mg ml–1 against E. coli, 8 mg ml–1
against B. cereus and 4 mg ml–1 against S. aureus (relative to 64, 16 and
8 mg ml–1, respectively, for D28).
Our linguistic approach to designing synthetic AmPs might be
successful because of the pronounced modular nature of natural
AmP amino-acid sequences. As we have shown, this approach can
be used to expand the AmP sequence space rationally without using
structure–activity information or complex simulations of the interactions of a peptide with a membrane. The peptides that we designed
are different from previously designed synthetic AmPs15 in that they
bear limited homology to any known protein, which might be desirable for AmPs used in clinical settings. Some researchers argue that
widespread clinical use of AmPs that are too similar to human AmPs
will inevitably elicit bacterial resistance, compromising our own natural defences and posing a threat to public health16. In addition, using
our approach to develop an arsenal of diverse antimicrobial agents
would further reduce concerns about the development of antimicrobial resistance. We hope that this approach will help to expand the
diversity of known AmPs well beyond those found in nature, possibly
leading to new candidates for AmP-based antibiotic therapeutics.
Our designed AmPs show some degree of homology with natural
AmPs because the grammars are based on native sequences. Peptide
D28, for example, was matched by grammars derived from 11 natural
AmPs, including brevinin, temporin and ponericin. However,
Smith-Waterman alignments of our designed peptides against all
natural AmPs in the Swiss-Prot/TrEMBL database reveal that the
degree of homology is, by design (see Methods), limited. In particular, our two most active peptides, D51 and D28, have only 50 and
60% sequence identity, respectively, with the nearest natural AmP.
Our linguistic design approach might be most valuable as a
method for rationally constraining a sequence-based search for
new AmPs. Diverse leads generated by our algorithms could be
optimized using approaches described in the literature17. But the
Table 1 | MICs of peptides against bacterial targets
8
16
32
64
128
Class
Designed
Shuffled
Natural
AmPs
E. coli
(gram-negative)
MIC
MIC
#256 mg ml–1 #64 mg ml–1
16/40
1/38
6/8
4/40
0/38
6/8
B. cereus
E. coli or B. cereus
(gram-positive)
MIC
MIC
MIC
#256 mg ml–1
#64 mg ml–1
#256 mg ml–1
8/40
2/38
4/8
4/40
1/38
3/8
18/40
2/38
6/8
Table 2 | MICs of peptides against S. aureus and B. anthracis
MIC (mg ml–1)
256
.256
S. aureus
B. anthracis
D28
D51
D28, D51
D22
D63, S51
D5, D35, D43
S5, S22, S28
S35, S43, S63
D, designed on the basis of motifs; S, shuffled control.
868
©2006 Nature Publishing Group
D22
D5, D35,
D43, D63
S51
S5, S22, S28,
S35, S43, S63
LETTERS
NATURE | Vol 443 | 19 October 2006
linguistic approach described here has a number of limitations. First,
sequence families that are poorly conserved on an amino-acid level
would not benefit from this approach. Second, we suspect that the
small size of AmPs is helpful. Owing to the simple nature of regular
grammars, they would be less useful for designing larger proteins
and, in particular, proteins with complex tertiary or quaternary
structures.
3.
4.
5.
6.
7.
8.
METHODS
Computational design of unnatural AmPs. To design unnatural AmPs, we
combinatorially enumerated all grammatical sequences based on the set of
,700 grammars (see Supplementary Materials, Methods, Derivation of grammars). First, for each grammar, we wrote out all possible grammatical aminoacid sequences. So, for example, the grammar [IVL]K[TEGDK]V[GA]K
[AELNH][VA][GA]K produced 600 sequences, where 3 3 5 3 2 3 5 3 2 3 2 5
600, owing to the option of choosing one of many amino acids at each bracketed
position. There are roughly 3 million such 10-mers that correspond to antimicrobial patterns. Then we wrote out all possible 20-amino-acid sequences for
which each window of 10 amino acids is found in the set of 3 million 10-mers.
From this set, we removed any 20-mers that had six or more amino acids in a row
in common with a naturally occurring AmP. There are roughly 12 million such
20-mers, each of which is a ‘tiling’ of ten 10-mers.
We clustered these 12 million sequences using the Mcd-hit software18 at 70%
identity. From these clusters, we chose 42 high scoring sequences to test experimentally (see Supplementary Materials, Methods, Scoring of sequences)19. These
sequences have varying degrees of similarity to naturally occurring AmPs, as
determined by sequence alignment and discussed in the text.
Peptide synthesis and MIC determination. Fmoc (fluorenylmethoxycarbonyl)
chemistry was used to synthesize peptides on the Intavis Multipep Synthesizer
(Intavis LLC) at the MIT Biopolymers Lab, and confirmed using MALDI-TOF
Mass Spectrometry and recombinantly produced AmP standards. The MIC was
measured with an assay based on the NCCLS M26A and the Hancock assay for
cationic peptides17. See Methods in Supplementary Materials for details on synthesis and antimicrobial assay.
Received 1 May; accepted 4 September 2006.
1.
2.
Zasloff, M. Antimicrobial peptides of multicellular organisms. Nature 415,
389–-395 (2002).
Hancock, R. E. & Patrzykat, A. Clinical development of cationic antimicrobial
peptides: from natural to novel antibiotics. Curr. Drug Targets Infect. Disord. 2,
79–-83 (2002).
9.
10.
11.
12.
13.
14.
15.
16.
17.
18.
19.
Tiozzo, E., Rocco, G., Tossi, A. & Romeo, D. Wide-spectrum antibiotic activity of
synthetic, amphipathic peptides. Biochem. Biophys. Res. Commun. 249, 202–-206
(1998).
Biragyn, A. et al. Toll-like receptor 4-dependent activation of dendritic cells by
b-defensin 2. Science 298, 1025–-1029 (2002).
Ellerby, H. M. et al. Anti-cancer activity of targeted pro-apoptotic peptides. Nature
Med. 5, 1032–-1038 (1999).
Giangaspero, A., Sandri, L. & Tossi, A. Amphipathic a helical antimicrobial
peptides. Eur. J. Biochem. 268, 5589–-5600 (2001).
Shai, Y. Mode of action of membrane active antimicrobial peptides. Biopolymers
66, 236–-248 (2002).
Jurafsky, D. & Martin, J. H. Speech and Language Processing: An Introduction to
Natural Language Processing, Computational Linguistics, and Speech Recognition
(Prentice Hall, Upper Saddle River, New Jersey, 2000).
Searls, D. B. The language of genes. Nature 420, 211–-217 (2002).
Hofmann, K., Bucher, P., Falquet, L. & Bairoch, A. The PROSITE database, its status
in 1999. Nucleic Acids Res. 27, 215–-219 (1999).
Rigoutsos, I. & Floratos, A. Combinatorial pattern discovery in biological
sequences: The TEIRESIAS algorithm. Bioinformatics 14, 55–-67 (1998).
Wang, Z. & Wang, G. APD: the Antimicrobial Peptide Database. Nucleic Acids Res.
32, D590–-D592 (2004).
Bairoch, A. & Apweiler, R. The SWISS-PROT protein sequence database and its
supplement TrEMBL in 2000. Nucleic Acids Res. 28, 45–-48 (2000).
Wu, M. & Hancock, R. E. W. Interaction of the cyclic antimicrobial cationic peptide
bactenecin with the outer and cytoplasmic membrane. J. Biol. Chem. 274, 29–-35
(1999).
Tossi, A., Sandri, L. & Giangaspero, A. Amphipathic, a-helical antimicrobial
peptides. Biopolymers 55, 4–-30 (2000).
Bell, G. & Gouyon, P-H. Arming the enemy: the evolution of resistance to selfproteins. Microbiology 149, 1367–-1375 (2003).
Hilpert, K., Volkmer-Engert, R., Walter, J. & Hancock, R. E. W. High-throughput
generation of small antibacterial peptides with improved activity. Nature
Biotechnol. 23, 1008–-1012 (2005).
Li, W., Jaroszewski, L. & Godzik, A. Clustering of highly homologous sequences to
reduce the size of large protein database. Bioinformatics 17, 282–-283 (2001).
Maizel, J. V. & Lenk, R. P. Enhanced graphic matrix analysis of nucleic acid and
protein sequences. Proc. Natl Acad. Sci. USA 78, 7665–-7669 (1981).
Supplementary Information is linked to the online version of the paper at
www.nature.com/nature.
Acknowledgements The authors would like to thank M. Zasloff, K. D. Wittrup,
R. Berwick, and G. Georgiou for valuable input on the draft manuscript, and
J. Moxley for figure preparation. The authors gratefully acknowledge the support of
the Singapore-MIT Alliance, the NIH, and the Fannie and John Hertz Foundation.
Author Information Reprints and permissions information is available at
www.nature.com/reprints. The authors declare no competing financial interests.
Correspondence and requests for materials should be addressed to G.S.
([email protected]).
869
©2006 Nature Publishing Group